International Journal of Scientific Research in Computer Science, Engineering and Information Technology

ISSN : 2456-3307 (www.ijsrcseit.com)

doi : https://doi.org/10.32628/CSEIT228630

Face Recognition Using Facenet Deep Learning Network for

Attendance System
Rutuja Bankar, Nikita Bargat, Isha Hanmante, Prof. Hemlata Dakore
Computer Science Engineering, G. H. Raisoni Institute of Engineering and Technology, Nagpur, Maharastra,
India
ABSTRACT
Face recognition that is technology used for recognizing human faces based on
Article Info certain patterns and re-detect faces in various conditions. Face recognition is
currently becoming popular to be applied in various ways, especially in security
Publication Issue : systems. Various methods of face recognition have been proposed in researches
Volume 8, Issue 6 and increased accuracy is the main goal in the development of face recognition
November-December-2022 methods. FaceNet is one of the new methods in face recognition technology.
This method is based on a deep convolutional network and triplet loss training
Page Number : 458-463 to carry out training data, but the training process requires complex computing
and a long time. By integrating the Tensorflow learning machine and pre-
Article History trained model, the training time needed is much shorter. This research aims to
Accepted: 20 Nov 2022 conduct surveys, test performance, and compare the accuracy of the results of
Published: 05 Dec 2022 recognizing the face of the FaceNet method with various other methods that
have been developed previously. Implementation of the FaceNet method in
research using two types of pre-trained models, namely CASIA-WebFace and
VGGFace2, and tested on various data sets of standard face images that have
been widely used before. From the results of this research experiment, FaceNet
showed excellent results and was superior to other methods. By using VGGFace2
pre-trained models, FaceNet is able to touch 100% accuracy on YALE, JAFFE,
AT & T datasets, Essex faces95, Essex grimace, 99.375% for Essex faces94 dataset
and the worst 77.67% for the faces96 dataset.
Keywords — Face Recognition, Face Detection, FaceNet, Deep Convolutional
Network, TensorFlow, Deep Learning

I. INTRODUCTION recognition methods in the last two decades shows

Face recognition is one branch of computer science is
an ability to recognize or identify the person's
identity by analyzing the pattern-based facial
contours of human faces. The development of face
very rapid progress. Initially, the method was running
very slowly with the results of poor accuracy so that
it cannot be applied daily life until now it can be
applied in real-time by producing excellent accuracy.
Currently, face recognition is used as a technology to

Copyright: © the author(s), publisher and licensee Technoscience Academy. This is an open-access article distributed 458
under the terms of the Creative Commons Attribution Non-Commercial License, which permits unrestricted non-
commercial use, distribution, and reproduction in any medium, provided the original work is properly cited
 Rutuja Bankar et al Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol., November-December-2022, 8 (6) : 458-463
provide multiple security in various practices like
verification of identity, access authority, observation,
to replace passwords and identity cards that are no
longer safe. The use of face recognition has the
benefit of verifying personal data because inhuman
faces things like irises, retinas, and faces are very
unique to each other.
Face Recognition has many methods in its application
today. For Ex: FaceNet, OpenFace, etc. FaceNet is one
of the uses of face recognition based on Deep
Learning. That is a one-shot learning method using
Euclidean space to calculate the similarity distance for
each face. FaceNet is a fairly new method, introduced
by Google researches in 2015, using Deep
Convolutional Network method. In the early 1990s,
traditional face recognition systems were not stable
and still did not appear deep learning and had several
errors in real time applications. Nowadays, deep
learning appears and it is especially good for
recognition and detection. Deep learning acts like a
human brain, learning by itself. In deep neural
network architecture, when we create a neural
network, the more the hidden layers (neurons), the
better the accuracy. In this paper, FaceNet is used for
extracting features from faces. Support vector
machine is used for classification. To train the model,
‘triplet loss’ is a loss function. The purpose of the
system is to easily know which students are in the
classroom in a short time by implementing an
automated attendance management system by using
face recognition techniques and to save time
Volume 8, Issue 6, November-December-2022 | http://ijsrcseit.com
consuming for taking attendance. The main purpose is
to develop a reliable system by using deep learning.
The system will record the attendance of the student
automatically by matching from the training data-set.
II. METHODOLOGY
FaceNet
FaceNet was built by Google researchers using a Deep
Convolutional Neural Network (DCNN) that maps
images of a person's face into Euclidean spaces
(collections of geometrical points) which are also
called embedding.
Embedding is obtained from the level of similarity
and differences in faces, so that if the face has a
similarity the value will get closer, and if the face is
different the value will get farther.
FaceNet is a method that uses deep convolutional
networks to optimize its embedding, compared to
using intermediate bottleneck layers as a test of
previous deep learning approaches. This method is
called one-shot learning. In more detail, this method
can use a small sample of face images to produce the
initial model, and when there are new models, the
initial model can be used without retraining. FaceNet
directly trains the face using the Euclidean space
where the distance consists of similarities between
facial models. When the results of similarities
between face models are obtained, it will be easy to
carry out face recognition and classification using
FaceNet attached become feature vectors.
In the training process, FaceNet applies triplets by
matching face to face with the online novel triplet
mining method. Of course, this triplet consists of a
collection of anchor images, where each image
consists of positive and negative images. Fig. 1 shows
the structural model used in FaceNet. FaceNet
consists of batch layers as input and deep architecture
which is deep CNN followed by L2 normalization,
that become the result of face embedding FaceNet
459
 Rutuja Bankar et al Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol., November-December-2022, 8 (6) : 458-463

also pursued by the triplet loss when the training
process.
Triplet loss training methods have three main
elements namely anchor, positive and negative. This
triplet loss works by minimizing the distance between
anchors positively and maximizing the distance
between anchors negatively. Where this positive has
the same identity as the anchor and negative has a
different identity from the anchor.
inputs for face detection and recognition and as well
creates multiple copies of the proposed images.
The second neural network, R-Net, is known as the
refine network. It feeds on the proposed image inputs
from P-Net and refines by filtering the false candidate
inputs from the true candidates.
The third and last neural network, O-Net, does the
rest of the job. It concludes the operation by finding
five facial landmarks among the candidate image
inputs. A facial landmark contains key features of the
human face.

FaceNet trains its output directly into concise 128-

dimensional embedding by apply triplet based loss
method depend on LMNN. It formed by two
thumbnails of compared faces and thumbnails that do
not match and the loss aim to distinguish between
Tensorflow
positive and negative pairs using a range of limit.
Using TensorFlow to build face recognition and
Thumbnails were cut tightly on the face field, it
detection models might require effort, but it is worth
didn’t need 2D or 3D adjustment, apart from the ratio
it in the end. As mentioned, TensorFlow is the most
and translation implemented.
used Deep Learning framework and it has pre-trained
models that easily help with image classification. In
Preprocessing
most cases, to generate a model means the
At the preprocessing stage, three processes were
classification of the images only needs to provide a
carried out, namely, detection, resizing, and cropping
similar image which is the positive image. The image
faces using the library's help from Multi-Task
is then trained and retrained through a process
Cascaded Convolutional Neural (MTCNN). Detection
known as anchoring or Transfer Learning. Years back,
was used to determine the position of the face in a
finding that model for training and retraining was
given image which was then realized in the form of a
difficult. Now, TensorFlow has simplified the process.
bounding box. Then implemented the cropping based
Thanks to its huge open-source community, no one
on the bounding box. After obtaining a face image
has to go through the task of generating a model once
would be scaled (resize) according to the input size of
another developer from the other end of the world
the model. The MTCNN algorithm works through the
had done it for all to use.
three neural networks (P-Net, R-Net, and O-Net)
connected to a single cascading neural network that is
Classification
unique to each one of them.
Face recognition classification in this system uses
SVM because it has good performance and is widely
The first neural network, P-Net, is known as the
used in face recognition . SVM works by giving
proposal network. It proposes the candidate image
dividing boundaries to 2 adjacent classes. Margin is
the closest point between Hyperplane and the closest

Volume 8, Issue 6, November-December-2022 | http://ijsrcseit.com 460

 Rutuja Bankar et al Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol., November-December-2022, 8 (6) : 458-463
point of each class which is then called the Support
Vector Machine Linear Support Vector Machine
(SVM) is very effective and widely used when
separating face embedding vectors.
By adjusting linear SVM to training data and setting
'kernel' attribute to 'linear'. In making predictions,
will use the probability used to set it to be true. In
classifying the data is divided into 80% training data
and 20% test data. The two groups of data will be
processed by training and save the SVM classification
model in a file. Classification performance will be
evaluated using a confusion matrix.
III. PROPOSED SYSTEM
The purpose of this work is to implement a system
that can automatically mark the attendance of the
students present in a classroom without the
involvement of the teacher. Here we are introducing
an automatic attendance marking system by face
recognition using FaceNet. The ultimate objective of
the system is to improve and organize the class
attendance system and to decrease the number of
errors occurred in the manual process and to provide
attendance reports periodically For this proposed
system, initially we have to collect the required
dataset from the students, which the photos of the
students and store them in the database. The students
must provide pictures based on the sample pictures.
The sample should have ten images from different
angles. These ten images of each student will be
stored in the database with their name/roll no as the
file name. These 10 images will be stored in a folder
with their name/roll no. The operations like blurring,
contrast/brightness, rotation will be applied for each
of the images. From the above mentioned ten images
eight images will be taken for the training purpose
and two images will be taken for the testing purpose.
The system can be divided into two phases testing
phase and prediction phase.
Volume 8, Issue 6, November-December-2022 | http://ijsrcseit.com
After collecting the required dataset, we pass it for
the face detection purpose. For face detection, here
we are using MTCNN (Multi-Task Convolutional
Neural Network). MTCNN does both face detection
and face alignment and the process consists of three
stages of convolutional networks that are able to
recognize faces and landmark location such as eyes,
nose and mouth. The MTCNN will detect the faces
from the given images and extract the features from
the faces and convert them into unique arrays. Then
those extracted features will be stored in separate files
for each students with their name as file name. Then
FaceNet model should be loaded into the system and
it will be trained with the extracted features of each
students. Get two images from each student’s dataset
for testing and test those images with the FaceNet
model. The FaceNet has an accuracy of about 99.63 %,
which makes this system more accurate. FaceNet
takes an image of the person’s face as input and
outputs a vector of 128 numbers which represent the
most important features of a face which is called as an
embedding. The FaceNet transforms the face image
into 128-dimensional vectors and place it in the
Euclidean space and the model thus created is trained
for triplet loss to capture the similarities and
differences on the image dataset provided. Finally the
SVM (Support Vector Machine Classifier) algorithm is
used for the classification of each trained model.
461
 Rutuja Bankar et al Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol., November-December-2022, 8 (6) : 458-463

Thus, it proved to be a more time-saving and secure

system than traditional attendance systems.
We can extend the work in order to implement a
fully-featured attendance system using facial
recognition and marking the attendance in order to
provide a more real-time analysis of attendance data.

V. REFERENCES

[1]. Edwin Jose, Greeshma M., Mithun Haridas T.

A camera will be placed in the classroom, which will P., Supriya M. H, “Face Recognition based
capture an image that containing all the students that Surveillance System Using FaceNet and
present in the classroom, in a regular interval. From MTCNN on Jetson TX2” in 2019 5th
the captured image, faces of each student will be International Conference on Advanced
extracted and then they are encoded with the Computing & Communication Systems
FaceNet model. The result will be passed to the (ICACCS).
trained model for comparison and determining the [2]. Mingjie He, Jie Zhang, Shiguang Shan, Meina
corresponding student. Then we get the names/roll.no Kan, Xilin Chen, “Deformable Face Net for Pose
of the corresponding students as the output for Invariant Face Recognition” in Pattern
marking attendance, which will help us to mark the Recognition.
attendance of the students. The recorded attendance [3]. Timmy Schenkel, Oliver Ringhage, Nicklas
will be stored in the database, which the teacher can Brandin,. “A COMPARATIVE STUDY OF
later access by logging in through the website which FACIAL RECOGNITION TECHNIQUES”
will be available asexcel sheet. Bachelor Degree Project in Information
Technology , Basic level 30 ECTS ,Spring term
2019.
[4]. Ghaith Bouallegue, Ridha Djemal, “EEG person
identification using Facenet, LSTM-RNN and
SVM” in 2020 17th International Multi-
Conference on Systems, Signals & Devices
(SSD'20).
[5]. Kyle Heath, Leonidas Guibas, “FACENET:
TRACKING PEOPLE AND ACQUIRING
CANONICAL FACE IMAGES IN A WIRELESS
IV. CONCLUSION CAMERA SENSOR NETWORK” .
[6]. Augusto F. S. Moura1, Silas S. L. Pereira1 ,
Face Recognition based Attendance System Using Mario W. L. Moreira ́, and Joel J. P. C.
FaceNet and MTCNN was successfully implemented. Rodrigues,”Video Monitoring System using
As a result of this standalone system, the person in Facial Recognition: A Facenet-based Approach”
the dataset was detected and an excel sheet of the in 2020 EEE.
current students was created with an accuracy of

Volume 8, Issue 6, November-December-2022 | http://ijsrcseit.com 462

 Rutuja Bankar et al Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol., November-December-2022, 8 (6) : 458-463

[7]. Thida Nyein, Aung Nway Oo, “University posteriori model adaptation. IEEE Transactions
Classroom Attendance System Using FaceNet on Pattern Analysis and Machine Intelligence,
and Support Vector Machine”. 29(4):743–748, April 2007.
[8]. Florian Schroff, Dmitry Kalenichenko, James
Philbin, “FaceNet: A Unified Embedding for Cite this article as :
Face Recognition and Clustering”, CVPR2015.
[9]. S. S. Thomas, S. Gupta and V. K. Subramanian, Rutuja Bankar, Nikita Bargat, Isha Hanmante, Prof.
"Smart surveillance based on video Hemlata Dakore, "Face Recognition Using Facenet
summarization," 2017 IEEE Region 10 Deep Learning Network for Attendance System",
Symposium (TENSYMP), Cochin, 2017, pp. 1-5. International Journal of Scientific Research in
doi: 10.1109/TENCONSpring.2017.8070003 Computer Science, Engineering and Information
[10]. Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li: Technology (IJSRCSEIT), ISSN : 2456-3307, Volume 8
“Joint Face Detection and Alignment using Issue 6, pp. 458-463, November-December 2022.
Multi-task Cascaded Convolutional Networks”, Available at doi :
2016; [http://arxiv.org/abs/1604.02878 https://doi.org/10.32628/CSEIT228630
arXiv:1604.02878]. DOI: Journal URL : https://ijsrcseit.com/CSEIT228630
[https://dx.doi.org/10.1109/LSP.2016.260334210
.1109/LSP.2016.2603342].
[11]. S. D. Shendre, “An Efficient way to Trace
Human by Implementing Face Recognition
Technique using TensorFlow and FaceNet API,”
International Journal for Research in Applied
Science and Engineering Technology, vol. 6, no.
4, pp. 605–608, 2018.
[12]. Z. He, M. Kan, J. Zhang, X. Chen, S. Shan, A
fully end-to-end cascaded cnn for facial
landmark detection, in: IEEE International
Conference on Automatic Face Gesture
Recognition (FG), 2017, pp. 200–207.
[13]. I. Masi, S. Rawls, G. Medioni, P. Natarajan,
Pose-aware face recognition in the wild, in:
IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), 2016, pp. 4838–
4846.
[14]. Gary B. Huang, Manu Ramesh, Tamara Berg,
and Erik Learned- Miller. Labeled Faces in the
Wild: A Database for Studying Face
Recognition in Unconstrained Environments.
University of Massachusetts, Amherst,
Technical Report 07-49, October, 2007.
[15]. S. Marcel, J. R. Mill ́an. Person authentication
using brainwaves (EEG) and maximum a

Volume 8, Issue 6, November-December-2022 | http://ijsrcseit.com 463